The Hack-a-saurus Guide to SAS® Grid: Saving Time with ......2020/01/29 · 1. SAS grid is not always available 2. SAS datasets have no mechanism for managing parallel dataset access

1

Paper 4949-2020

The Hack-a-saurus Guide to SAS® Grid: Saving Time with

Parallel Database Queries

Ted D. Williams, PharmD, BCPS, Magellan Method

ABSTRACT

Have you ever worked overtime or missed a deadline because it took hours (even days) to

extract the data you needed from multiple sources? And heaven help us if there is a flaw in

the query, a change in data structures, new customer requirements, or a execution error

after 12 hours of runtime. But a new dawn has broken; SAS® Grid supports parallel

processing. SAS Grid enables a single program to query multiple data sources in parallel

making retrieval and aggregation as simple as F3. Unlike other parallel processing solutions,

SAS Grid does not require learning a completely new syntax (I'm looking at you, DS2).

Unfortunately, it is not a simple as standing up a SAS Grid server and running Enterprise

Guide ‘Analyze for Grid’. Do you live in the WORK library? Sorry, no longer shared. Like

global variables? Too bad, those break too. Do you use macros? Tough luck, SAS Grid

doesn’t know about your complied macros. But do not despair! All these issues can be

addressed to without massive rewriting. This paper will save you those tears and overtime

hours by giving you the solutions to these problems. Reading this paper is like getting into a

prestigious university without your parents having to bribe administrators!

INTRODUCTION

Extraction, Transformation, and Loading (ETL) of data can be one of the most time-intensive

tasks for a data analysis project. This is particularly true for projects involving multiple data

sources that need to be merged or stacked together. Often retrieving each dataset is an

independent task. Because SAS can only execute one command at a time, the data

extraction must be run sequentially, one after the other. Multiple SAS instances can be

executed manually to collect the data, but babysitting these processes is a poor use of

human capital and susceptible to mistakes. A fully automated parallel process is possible

without completely rewriting your code using SAS Grid. This paper discusses a few tips and

tricks for the minor rewriting and troubleshooting required.

SERIAL PROCESSING OF INDEPENDENT TASKS

Traditional ETL with SAS involved sequential steps, such as:

*A macro library with DataSource macros;

%INCLUDE '\\MyNetwork\SASGF_2020_LIB.sas';

%MACRO GetSerialData(Results=);

PROC SQL;

create table &Results. (Source CHAR(100),Value CHAR(1));

QUIT;

%DataSourceOne(Results=&Results.);

%DataSourceTwo(Results=&Results.);

%DataSourceThree(Results=&Results.);

%mend;

%GetSerialData(Results=WORK.Serial);

2

The macro %GetSerialData creates a dataset then calls a series of macros, stacking the

data each data source into a single dataset (WORK.Serial). Each of the data source macros

are independent, except for the shared destination. It is the independent nature of these

tasks which makes them ideal for parallelization.

GRID BASICS

The mechanism for spawning parallel processes (aka remote submissions) using SAS Grid

can be condensed to:

%let RC = %SYSFUNC(grdsvc_enable(_all_,Server=SASApp));

OPTIONS AUTOSIGNON;

RSUBMIT MySes WAIT=NO;

%put NOTE: Server Host: &SysHostName.;

ENDRSUBMIT;

WaitFor _all_ ;

Signoff _all_;

The command grdsvc_enable enables the grid service and returns a value indicating the

status of the grid.1 A return code of zero indicates the grid has been enabled and is ready

for a new remote submission. Other return codes indicate various errors. OPTIONS

AUTOSIGNON allows SAS to automatically sign on to the grid server when a job is submitted.2

A manual sign on is possible and allows for greater control of the remote submission. The

remote submission is bound by the RSUBMIT and ENDRSUBMIT statements. At a minimum,

RSUBMIT requires the name of the remote submission (limited to 8 characters). WAIT=NO

indicates the parent process can continue while the SAS server execute the submission (i.e.,

runs in parallel).3 The WaitFor _all_ statement is where SAS execution stops until all

remote sessions has been completed.2 Finally, Signoff _all_ closes all the connections to

the grid server.

If this code was all that it took to run SAS in parallel in a production environment there

would be no need for this paper. There are a number of challenges that must be addressed

in order to transform %GetSerialData into %GetParallelData:

1. SAS grid is not always available

2. SAS datasets have no mechanism for managing parallel dataset access

3. The WORK library is not shared with spawned sessions

4. Complied macros are not available to spawned sessions

5. Macro variables are not shared with spawned sessions

OVERCOMING SAS® GRID CHALLENGES

MAKING SURE THE GRID IS AVAILABLE

In the real world, Grid servers are not always available for parallel processing. But that does

not mean your program is doomed to fail. If the Grid is not available, simply use the

traditional serial approach:

3

%MACRO GetParallelData(Results=);

PROC SQL;


QUIT;

%IF %SYSFUNC(grdsvc_enable(_all_,Server=SASApp)) NE 0 %THEN

%DO;

%PUT WARNING: no grid, no worries, run macros in series;




%END;

%ELSE

%DO;

/*More to come...*/

%END;

%MEND;

%GetParallelData(Results=WORK.Parallel);

MANAGING PARALLEL SESSIONS

SAS does not like multiple processes working on a single dataset. “Does not like” means

SAS will generate an error and fail. In this example, all the ETL macros update a single

dataset. To prevent concurrency errors, a dedicated temporary results dataset must be

created for each remote submission. After all remote submissions have executed the

temporary datasets can be stacked together without conflicts.

But where will the temporary datasets reside? Unfortunately, the usual destination, the

WORK library, is NOT shared among remote submissions. Simply using a local drive (e.g.

C:\ ) will not work since most GRID environments will have multiple physical servers. A

shared network location which is visible to all grid servers can be used. If the same ETL

macro is being run simultaneously by multiple clients, the issue of simultaneous dataset

edits can also occur on a shared network location. Addressing these concurrency issues

requires creating a unique, accessible location for each remote submission and mapping it

to a LibName. This functionality can easily be wrapped into a macro:

%SafeSpace(NewLib=, ParentLib= );

The parameter ParentLib should contain a existing LibName mapped to a shared network

location, such as:

LIBNAME Sandbox '\\MyNetwork\';

The parameter NewLib will be used to create a new, unique network location and SAS

library. The details of the SafeSpace macro declaration are included in the appendix. With a

dedicated library location, a temporary dataset can be created for each remote submission.

These individual temporary datasets will be merged after all parallel processing is

completed.

Using this concurrency management strategy, the GetParallelData macro definition

becomes:

4

%MACRO GetParallelData(WorkingLib=, Results= );

PROC SQL;


QUIT;


%DO;

%PUT WARNING: no grid, no worries, run macros in series;




%END;

%ELSE

%DO;

%PUT WARNING: Grid Available, run macros in parallel;

*Make sure the WORK-ing directory us usable by grid;

%IF %UPCASE(&WorkingLib.) = WORK %THEN

%DO;

%PUT ERROR: WORK is not shared between RSUBMITs.;

%GOTO Finish;

%END;

*Create a temp library for each remote submission;

%SafeSpace(NewLib=Temp1,ParentLib=&WorkingLib.);



*Make blank copies Results in each temp library;

PROC SQL;

CREATE TABLE Temp1.Results AS SELECT * FROM &Results.;



QUIT;

/*More to come...*/

%END;

%Finish:

%MEND;


%GetParallelData(WorkingLib=SANDBOX, Results=WORK.Parallel);

SHARING OF LIBRARIES, MACROS, AND MACRO VARIABLES ACROSS SESSIONS

No variables, libraries, or macro definitions are automatically passed to remote submissions.

Think of it as a completely new instance of SAS Enterprise Guide. Fortunately, there are

mechanisms to resolve these issues.

%SYSLPUT can be used to pass variables to the remote session.4 The basic syntax is:

%SYSLPUT <variables to pass> / REMOTE = <Submission Name>;

The first parameter <variables to pass> supports several options including _USER_,

_GLOBAL_ and _LOCAL_. _USER_ can produce some odd results when using nested macro.

5

_LOCAL_ transfers only locally defined macro variables and makes debugging much easier.

REMOTE =requires the name of the remote submission, as listed in the RSUBMIT statement.

SAS does not provide any mechanism for transferring complied macros to a remote

submissions. The %INCLUDE statement can be used to execute a SAS program in the remote

submission.6 Therefore, any macros the remote submission needs to run must be contained

within an included file. For this example, the macro definitions for %DataSourceOne,

%DataSourceOne, %DataSourceOne are all contained in the

\\MyNetwork\SASGF_2020_LIB.sas SAS source file. After executing the %INCLUDE

statement within the RSUBMIT block, the macros are available in the remote session.

The INHERITLIB =(<library name>) clause of the RSUBMIT statement transfers a LIBNAME

definition to the remote session.5 Beware that using INHERITLIB with a LIBNAME mapped

to an OLEDB connection will clear the parent LIBNAME. This can break a lot of downstream

code. LIBNAMEs mapped to file system locations do not have this issue. This problem can be

corrected by the SAS system administrators add the NOCONNECTEVENTS option to the

workspace server’s usermod config file. To avoid these issues, a LIBNAME statement can be

used within the RSUBMIT block, similar to the %INCLUDE discussed above.

Putting all of these techniques together:

OPTIONS AUTOSIGNON;

%SYSLPUT _LOCAL_ / REMOTE = Ses1;

RSUBMIT Ses1 WAIT=NO INHERITLIB=(Temp1);


%DataSourceOne(Results=Temp1.Results);

ENDRSUBMIT;


RSUBMIT Ses2 WAIT=NO INHERITLIB=(Temp2) ;


%DataSourceTwo(Results=Temp2.Results);

ENDRSUBMIT;




%DataSourceThree(Results=Temp3.Results);

ENDRSUBMIT;

MERGING PARALLEL RESULTS

The final steps are to wait for all parallel threads to finish, closing all severs sessions, and

merging the temporary files:

WaitFor _all_ ;

Signoff _all_;

DATA &Results.;

SET Temp1.Results

Temp2.Results

Temp3.Results

;

RUN;

6

CONCLUSION

Converting existing code to parallel processing using SAS Grid is not free, but can

dramatically reduce processing time and be a worthwhile investment. Enabling parallel

processing on the Grid (grdsvc_enable) and spawning remote sessions (RSUBMIT) are

straightforward. Passing variables and libraries to remote sessions are more complex but

are possible using intrinsic SAS features (%SYSLPUT, INHERITLIB). Executing macros and

merging datasets without causing errors are possible but requires minor rewriting. Because

of these complexities, using SAS grid should be consider in cases when the parallel

processes are independent and can be easily synchronized.

APPENDIX

SASGF_2020_LIB.SAS

%Macro IsNumeric(VarToCheck);

%local Result;

%let Result = %eval(not %sysfunc(verify(&VarToCheck.,-0123456789)));

&Result

%mend;

%Macro LibraryExist(lb);

%local Result;

%Let Result=-1;

%IF %LENGTH(&lb)=0 %Then

%Let Result=0;

%else

%do;

%let Result = %sysfunc(LibRef(&lb));

%if &Result = 0 %then

%let Result = 1;

%Else

%let Result = 0;

%end;

&Result %mend;

%macro SafeSpace(ParentLib=, NewLib=);

%local ParentPath;

%Local CreatedPath;

%local UniqueName;

%local Leader;

%let Leader=;

%let UniqueName =;

%let ParentPath=;

%LET CreatedPath=;

/*Default to work if no library was provided*/

%IF %LENGTH(&ParentLib.)=0 %THEN

%LET ParentLib = WORK;

%if %SUBSTR(%IsNumeric(&NewLib.),1,1) = 1 %then

%do;

%PUT ERROR: NewLib cannot start with a numeric character *&LibName.*;

%goto Finish;

%end;

7

%IF %length(&NewLib.)>8 %THEN

%do;

%PUT ERROR: NewLib must be <=8 characters &NewLib.*=%length(&NewLib.);

%goto Finish;

%end;

*Get the network path for the Existing Parent Library;

proc sql noprint;

SELECT DISTINCT path

INTO :ParentPath

FROM sashelp.vLibNam

WHERE UPCASE(libname)=UPCASE("&ParentLib.")

;

quit;

%If %Length(&ParentPath.)=0 %THEN

%do;

%PUT ERROR: Cannot find path for *&ParentLib.* in sashelp.vLibNam;

%Goto Finish;

%end;

*Create both a new file sub directory and a new library in sas

with a unique locations (as defined by macro %GetUniqueName();

*Get a unique name based on the server, calling macro and a datestamp;

%let UniqueName =UPCASE(&SYSHOSTNAME.)_%SUBSTR(%Sysfunc(datetime()),1,10);

*Limit length to 32;

%let Result=%Right(&UniqueName.,32);

*If name starts with a number, replace with a letter;

%let Leader = %substr(&UniqueName.,1,1);

%if %IsNumeric(&Leader.) = 1 %THEN

%do;

%let UniqueName = %SUBSTR(&UniqueName.,2);

%LET UniqueName=A&UniqueName.;

%end;

%let CreatedPath = %SYSFUNC(TRIM(&ParentPath.))\&UniqueName.&NewLib.;

options dlcreatedir;

libname &NewLib. "&CreatedPath.";

options NODLCREATEDIR;

*Just in case this file system folder was used before,

destroy anything in that folder.;

PROC DATASETS LIB=&NewLib. NOLIST NOWARN KILL;

RUN;

%Finish:

%mend;

%PUT Data Access Simulation Macros;

%MACRO DataSourceOne(Results=);

*Some dummy data, replace with database queries;

proc sql;

Insert into &Results. VALUES("&SysMacroName.",'A');

Insert into &Results. VALUES("&SysMacroName.",'B');

Insert into &Results. VALUES("&SysMacroName.",'C');

8

Insert into &Results. VALUES("&SysMacroName.",'D');

Insert into &Results. VALUES("&SysMacroName.",'E');

quit;

%MEND;

%MACRO DataSourceTwo(Results=);

proc sql;






quit;

%MEND;

%MACRO DataSourceThree(Results=);

proc sql;






quit;

%MEND;

%put NOTE: Done loading macro library;

SASGF_2020.SAS

*A macro library;


options mprint;

%MACRO GetSerialData(Results=);

PROC SQL;

create table &Results. (Source CHAR(100),Value CHAR(1)) ;

QUIT;




%mend;

%MACRO GetParallelData(WorkingLib=,Results=);

*CalledBy used to create a unique temporary WORK-ing library;

%LOCAL CalledBy;

%LET CalledBy=&SysMacroName.;

*Make sure the WORK-ing directory provided;

%IF %LibraryExist(&WorkingLib.)=0 %THEN

%DO;

%PUT ERROR: WorkingLib *&WorkingLib.* does not exist.;

9

%GOTO Finish;

%END;

*Create the final Results dataset;

PROC SQL;

create table &Results. (Source CHAR(100),Value CHAR(1)) ;

QUIT;

*Check Grid Status (zero is operational );


%DO;

%PUT WARNING: no grid, no worries, run macros in serial;




%END;

%ELSE

%DO;

%PUT WARNING: Grid Available, run macros in parallel;

*Make sure the WORK-ing directory us usable by grid ;

%IF %UPCASE(&WorkingLib.) = WORK %THEN

%DO;

%PUT ERROR: Grid servers cannot use WORK.;

%goto Finish;

%END;




*Make blank copies Results in each temp library;

PROC SQL;

CREATE TABLE Temp1.Results AS SELECT * FROM &Results. ;



QUIT;

*Use automatic sign on for simplicity;

OPTIONS AUTOSIGNON;

************************

Create First Grid Session: Ses1

***********************;

*Use SYSLPUT Inherit variables into Ses1;


RSUBMIT Ses1

WAIT=NO /*enables parallel processing*/

INHERITLIB=(Temp1)/*specifies SAS libraries to inherit*/

;

*Between RSUBMIT and ENDSUBMIT do everything you

would do for a new program.;

10

*include reference to macro library;


*Run the macro;

%DataSourceOne(Results=Temp1.Results);

ENDRSUBMIT;

************

Repeat for session 2

*************;




%DataSourceTwo(Results=Temp2.Results);

%put NOTE: Server Host: &syshostname.;

ENDRSUBMIT;

************

Repeat for session 3

*************;




%DataSourceThree(Results=Temp3.Results);

%put NOTE: Server Host: &syshostname.;

ENDRSUBMIT;

*Syncronize parallel processes;

WaitFor _all_ ;

*Sign off to free up grid resources;

Signoff _all_;

*Now merge parallel results without fear of locking;

DATA &Results.;

SET Temp1.Results

Temp2.Results

Temp3.Results

;

RUN;

*A bit of clean up;

PROC DATASETS LIB=Temp1 NOLIST NOWARN KILL; RUN;



%END;

%Finish:

%MEND;

%PUT;

%GetSerialData(Results=WORK.Serial);

11

Title Serial Results;

PROC SQL OUTOBS = 100;

SELECT * FROM Serial;

QUIT;


*

Note that the only difference between the calls is the

definition of a local variable for a new WORK-ing library

;

%GetParallelData(Results=WORK.Parallel,WorkingLib=SANDBOX);

Title Parallel Results;

PROC SQL OUTOBS = 100;

SELECT * FROM Parallel;

QUIT;

REFERENCES

1. SAS Help Center: GRDSVC_ENABLE Function. https://documentation.sas.com/?docsetId=gridref&docsetTarget=n1x4fhd14au95fn1hhcknbvtpn1b.htm&docsetVersion=9.4&locale=en. Accessed January 29, 2020.

2. Quick-Start to SAS(R) Programming on the Grid. Presented at the: SAS Global Forum 2019; March 25, 2019; Dallas, Tx.

3. Jacobsen G, Lavery R. A Parallel Processing Primer. Presented at the: NorthEast SAS Users Group; 2017.

4. %SYSLPUT Statement - 9.3. http://support.sas.com/documentation/cdl/en/connref/63066/HTML/default/viewer.htm#n1bc4zi0lp02ivn1oxn9gc1bsaha.htm. Accessed January 29, 2020.

5. Syntax for the RSUBMIT Statement and Command: RSUBMIT Statement and Command. http://support.sas.com/documentation/cdl/en/connref/61908/HTML/default/viewer.htm#a002590433.htm. Accessed January 29, 2020.

6. SAS Help Center: %INCLUDE Statement. https://documentation.sas.com/?docsetId=lestmtsref&docsetTarget=p1s3uhhqtscz2sn1otiatbovfn1t.htm&docsetVersion=1.0&locale=en. Accessed January 29, 2020.

ACKNOWLEDGMENTS

I would like to thank all of the Magellan Health SAS users who supported me while I refined

these techniques. Special thanks go out to Scott Smith at Magellan Health who answered

my incessant questions and without who’s help these technique would have been

unachievable. Additionally, I would like to thank XXX, XXX, XXX, and XXX for their editorial

assistance.

CONTACT INFORMATION

12

Your comments and questions are valued and encouraged. Contact the author at:

Ted D. Williams, PharmD, BCPS

Senior Clinical Program Manager / Health Economics & Outcomes Research (HEOR)

Magellan Health / Magellan Method

[email protected]

SAS and all other SAS Institute Inc. product or service names are registered trademarks or

trademarks of SAS Institute Inc. in the USA and other countries. ® indicates USA

registration.

Other brand and product names are trademarks of their respective companies.

Documents

The Hack-a-saurus Guide to SAS® Grid: Saving Time with ......2020/01/29 · 1. SAS grid is not always available 2. SAS datasets have no mechanism for managing parallel dataset access